Training procedure for using experimental vibrational spectra

Develop a rigorous training methodology to fine-tune E(3)-equivariant interatomic potential energy models using experimental vibrational spectroscopy data, including infrared, Raman, and inelastic neutron scattering spectra; specify how to incorporate symmetry selection rules, matrix element effects, an appropriate training schedule, and a principled loss function for comparing frequency spectra to facilitate learning.

Background

The authors demonstrate fine-tuning on DFT-simulated molecular vibrational spectra (benzene) and observe improved predictions for related molecules (toluene, phenol), suggesting that spectral information can correct and enhance the underlying energy model.

They propose extending this approach to experimental spectra but emphasize unresolved methodological issues, including handling symmetry selection rules, matrix element effects, training schedules, and the design of a suitable loss for spectral comparison, which collectively define an open challenge for robust training on real experimental data.